Printing out average time of kernel calls

Author: michael boulton

kernel_files/tea_leaf_cg_cl.cl

@@ -87,6 +87,7 @@ __kernel void tea_leaf_cg_solve_calc_ur
  __global       double * __restrict const rrn)
 {
     __kernel_indexes;
+
     __local double rrn_shared[BLOCK_SZ];
     rrn_shared[lid] = 0.0;
 

kernel_files/tea_leaf_ppcg_cl.cl

@@ -110,7 +110,7 @@ __kernel void tea_leaf_ppcg_solve_calc_sd
             sd[THARR2D(0, 0, 0)] = alpha[step]*sd[THARR2D(0, 0, 0)]
                                 + beta[step]*r[THARR2D(0, 0, 0)]*Mi[THARR2D(0, 0, 0)];
         }
-        else
+        else if (PRECONDITIONER == TL_PREC_NONE)
         {
             sd[THARR2D(0, 0, 0)] = alpha[step]*sd[THARR2D(0, 0, 0)]
                                 + beta[step]*r[THARR2D(0, 0, 0)];

ocl_errors.cpp

@@ -215,10 +215,14 @@ CloverChunk::~CloverChunk
     {
         fprintf(stdout, "@@@@@ PROFILING @@@@@\n");
 
-        for (std::map<std::string, double>::iterator ii = kernel_times.begin();
-            ii != kernel_times.end(); ii++)
+        std::map<std::string, double>::iterator ii = kernel_times.begin();
+        std::map<std::string, int>::iterator jj = kernel_calls.begin();
+
+        for (ii = kernel_times.begin(), jj = kernel_calls.begin();
+            ii != kernel_times.end(); ii++, jj++)
         {
-            fprintf(stdout, "%30s : %.3f\n", (*ii).first.c_str(), (*ii).second);
+            fprintf(stdout, "%30s : %10.3f ms (%.2f μs avg.)\n",
+                ii->first.c_str(), ii->second, 1e3*ii->second/jj->second);
         }
     }
 }